Gemini Diffusion vs UI-TARS-72B-DPO

Gemini Diffusion

Google DeepMind

Gemini Diffusion is an experimental text and code generation model from Google DeepMind, announced at Google I/O in May 2025 as the first diffusion-based language model to achieve quality comparable to autoregressive models on standard benchmarks. Unlike transformer-based models that predict tokens sequentially left-to-right, it generates entire blocks of text by iteratively refining noise — the paradigm used in image and video generation models — enabling faster sampling speeds and stronger mid-generation error correction for code and mathematical editing tasks. At announcement it was available only as an experimental demo via waitlist, with no public API, marking it as a research milestone rather than a production deployment.

ByteDance

UI-TARS-72B-DPO, released by ByteDance in early 2025, is a 72 billion parameter multimodal large language model from the UI-TARS family, built on Qwen-2-VL and fine-tuned for automated GUI interaction and computer control. It features native understanding of screenshots, UI elements, and web interfaces, achieving strong results across GUI benchmarks for perception, grounding, and agentic control. UI-TARS-72B-DPO targets computer-use agents, web automation, and applications requiring robust visual UI reasoning.